Anthropic, a ChatGPT competitor, announced this week that it disrupted a large-scale hacking campaign that it attributes to a Chinese state-sponsored group. The campaign, discovered in mid-September, used AI to automate much of the operation.
According to the company, the attackers manipulated Claude’s “agentic” capabilities—getting the system to act autonomously rather than simply respond to commands—to break into nearly 30 global targets. The victims included technology firms, financial institutions, chemical manufacturers and government agencies.
Anthropic said the hacking group used so-called “jailbreaking” techniques on Claude that allowed it to bypass internal safeguards. Once inside, the AI executed phishing, network infiltration and data extraction with limited human oversight.
While the company detected and halted the campaign early, it views it as a watershed moment for cybersecurity. The idea that an AI system can run large parts of a cyber-espionage operation marks a sharp shift in how threats might evolve.
This event raises urgent questions for every business that uses AI agents or cloud-based models. If one system can be bent into attack mode, many more could follow—and the automated scale of such threats could outpace traditional defenses.